diff --git a/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh b/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh
index 30149448..f48cd0ce 100755
--- a/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh
+++ b/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh
@@ -75,6 +75,7 @@ COMMON_FLAGS=(
     # CUB needs this for non-NVCC compilers
     -D__CUDA_ARCH_LIST__=700
     -DCUB_DISABLE_NAMESPACE_MAGIC
+    -DCUB_WRAPPED_NAMESPACE=cccl_preload
 )
 
 if [[ "${CXX}" == *clang++* ]]; then
diff --git a/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu b/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu
index 50061632..47b33448 100644
--- a/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu
+++ b/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu
@@ -41,8 +41,8 @@ static constexpr unsigned int ALLOC_MAX_BIN      = 32;  /* 4 GB */
 static constexpr size_t       ALLOC_MAX_CACHED   = (size_t)8 * 1024 * 1024 * 1024; /* 8GB */
 
 /* ---- Global allocator singleton ---- */
-static cub::CachingDeviceAllocator& get_allocator() {
-    static cub::CachingDeviceAllocator instance(
+static cccl_preload::cub::CachingDeviceAllocator& get_allocator() {
+    static cccl_preload::cub::CachingDeviceAllocator instance(
         ALLOC_BIN_GROWTH,
         ALLOC_MIN_BIN,
         ALLOC_MAX_BIN,